Trust-aware denoising autoencoder with spatial-temporal activity for cross-domain personalized recommendations
Introduction
The evolution of Internet technologies has led to the development of a large number of online e-commerce and social-networking applications. In the past couple of years, the volumes of data have increased dramatically due to the onset of COVID-19 pandemic that motivated a large number of people to go for online purchases. This has posed new challenges for information retrieval systems that not only need to search the required content from diversified sources but also the content needs to be most relevant to an individual’s preferences.
Recommender systems have been adopted as viable solutions to address information overload problem as their aim is to suggest only relevant items to a user matching with an individual’s preferences.
Despite significant progress in collaborative filtering-based recommendation systems, there are still open challenges related to data sparseness and cold start that affect the performance of recommender systems. Data sparsity occurs when there are fewer entries provided by users for each item in user-to-item matrix. This results into large number of zero similarity values, leading to irrelevant recommendations for a user. Cold start problem arises when a user is new to the system and the system does not have a sufficient records for computing future recommendations.
Numerous methods have been devised in the past to address the aforementioned challenges. One approach is to embed the trust metrics in recommendation models that helps alleviating the drawbacks of cold start problem. According to sociology, trust is a set of expectations shared by all those involved in the interactions and/or networks [59]. Several trust-aware recommendation systems have been developed recently that generate more specific recommendations for a user depending upon the users’ context such as time, location, and current trust level. A users’ context may be static or dynamic [63]. For instance, trust or location of a user may vary with time. When a user’s context is encoded in the item- and user-specific latent factors, this helps a recommender system to solve the key issues, such as sparsity and cold start [76].
Numerous deep learning-based recommendation architectures have been proposed recently to address the data sparsity and cold start problems in traditional recommender systems [22], [67]. Deep learning effectively captures the non-linear relationships between users and items and provides more complex data abstractions in higher layers [24]. Moreover, from rich sources of data such as textual, and visual information, deep neural networks effectively capture the complex relationships within the data itself [41], [75]. Recommender systems, using cross-domain information, present a new dimension for solving cold start problem by shifting knowledge from source to target domains. One of the advantages of cross-domain systems is that information can be learned from dense domain to sparse domain, on the basis of a similarity score that is computed based on features of items or users [32].
Several works, such as [16], [20], [67], [36] adopted the collaborative filtering approaches for solving sparsity and cold-start problem in recommendation systems. Collaborative filtering (CF) uses similarity values computed based on users' preferences in the form of explicit feedback, usually rated from 1 to 5 scale [65], [31]. However, some users are likely to express their opinion in the form of text. Reviews in the form of text provide better guidelines for any new user to choose or get to know about items’ information.
One of the predominant challenges in e-commerce is the textual feedback biasness that is caused by personal interests of users, current trends, and the choice of users in terms of social influence. Due to this biasness, the performance of recommender systems is affected in terms of accuracy. For instance, Fig. 1, shows the biasness in the reviews of AliExpress1 given by different users belonging to ‘Russia (RU)’ in the domain of ‘Apparel Accessories’. It can be observed that both users have given 5-star ratings where first user’s review is positive (A***P) and second user’s review is negative (V***r), thus introducing biasness. Sometimes, rating systems are targeted by spammers who falsely modify the ratings of certain products [29].
Most of the existing CF systems ignore the role of users’ textual feedbacks’ despite that these having significant importance in overall product rating, especially when enough numerical ratings are not available [31]. So, it is necessary to bridge the gap between recommendations generated based on ratings and reviews entered as free text, to reach the unbiased rating of items.
In recommender systems, trustworthiness is one of the major issues, for example, malicious activity can change the preferences information that affects the recommendation accuracy [14]. When users want to buy products, then they may have more trust in the feedback provided by their friends (neighbors) as compared to ordinary users. Since the accuracy of rating prediction depends upon the neighbors, therefore, selection of best neighbors who have similar interests in cross-domain, is a challenge [78].
Moreover, recommendation time is another important factor not usually considered by existing literature. It defines the instant when a user may show interest in a particular item. To increase their sales, retailers may want to observe the purchase patterns of certain group of buyers with respect to time. This makes it a challenging problem for most of the existing recommender systems that do not consider the time factor at a finer granularity (e.g., hours, days, weeks, etc.) [68], [36]. By mining a user’s time-variant purchase patterns, we can predict a user's spatial–temporal preference activity. For example, Charlie is interested in buying apparel accessories, so a recommendation of apparel accessories in a nearby POI at a specific time would be convincing.
To address the above-mentioned challenges, our motivation is to provide an effective approach with maximum recommendation accuracy for generating trust-aware recommendations by assuming the scenario of ‘Users Overlap’ in cross-domain e-commerce systems. We have introduced the concept of preference bias which means that users have matching preferences in the regions that are frequently visited by the same users. The assumption is that an improved top-N recommendation can be generated if the users have a strong preference bias in their frequented regions and may have more trust in each other if they belong to the same region.
The following is the summarized list of our contributions:
- •
We propose a recommendation model called Trust-Aware Spatial-Temporal Activity based Denoising Autoencoder (TSTDAE) that models nonlinear relationships based on autoencoder in cross-domain scenario of ‘User Overlap’ for solving a cold start problem.
- •
To address feedback bias and improve the accuracy of recommendations, we perform sentiment analysis and filter out the biased users’ preferences based on negative polarity scores.
- •
For generating features, autoencoder is used by considering the unbiased users’ profiles and then performing clustering for finding neighbors for an active user to whom top-N items are recommended.
- •
The concept of trust degree and ratio of preference bias is introduced between trustors and trustees and the system ensembles this trust degree and ratio of preference bias with Denoising autoencoder for improving the recommendation quality.
- •
We project the best time of recommendations for an individual user or group of users using clustering on time series data and compute the preferable time of users’ purchases.
- •
We evaluate the performance of TSTDAE on public dataset of ‘aliexpress.com’ using Precision, MAP, NDCG, and HR.
The rest of the paper is organized as follows. Section 2 describes the related work. In section 3 we discuss the proposed model. Section 4 explains the experimental setup, Section 5 describes evaluation and results and Section 6 discusses the conclusions and future work.
Section snippets
Related work
In this section, we discuss a few of the most relevant works on cross-domain, trust-aware, and neural network-based recommendations. Table 1 presents a summary of state-of-the-art approaches related to cross-domain recommendations.
Proposed model
The architecture of the proposed system is shown in Fig. 2. The TSTDAE model consists of the six major modules: (a).
Experimental setup
In this section, we discuss about a dataset, domain partitioning, and data sampling.
Evaluation and results
In this section, we present the evaluation results of the proposed model compared to the baselines.
Conclusions and future work
E-commerce-based recommender systems face numerous challenges such as data sparsity, falsified feedback, and cold start problem. The most relevant item selection and its recommendation at the right time are the key challenges in most e-commerce recommender systems. Recently, deep learning and trust-based recommender systems have been the primary focus to solve the cold start problem and to bring intelligence to e-commerce sites for increasing productivity. In this paper, we proposed a model
CRediT authorship contribution statement
Adeel Ahmed: Conceptualization, Formal analysis, Investigation, Methodology, Validation, Writing – original draft, Experiments, Dataset Creation, and Validation. Khalid Saleem: Conceptualization, Investigation, Supervision, Writing – review & editing. Osman Khalid: Writing – review & editing. Jiechao Gao: Writing – review & editing, Resources, manuscript and performed formal analysis. Umer Rashid: Writing – review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Adeel Ahmed received his MPhil degree in computer science from Quaid-i-Azam University, Islamabad, Pakistan, in 2011. He is now PhD student at department of Computer Science, Quaid-i-Azam University, Islamabad, Pakistan. He was with the software industry for some years. His research interests include Machine learning, Recommendation systems, Social Network Analysis and Information Visualization.
References (78)
- et al.
On deep neural network for trust aware cross domain recommendations in E-commerce
Expert Syst. Appl.
(2021) - et al.
An intelligent recommender system using social trust path for recommendations in web-based social networks
Procedia Comput. Sci.
(2019) - et al.
A practical tutorial on autoencoders for nonlinear feature fusion: Taxonomy, models, software and guidelines
Inf. Fusion
(2018) - et al.
Social network-based service recommendation with trust enhancement
Expert Syst. Appl.
(2014) - et al.
Hybrid PoS-tagging: A cooperation of evolutionary and statistical approaches
Appl. Math. Model.
(2014) - et al.
An oppositional-Cauchy based GSK evolutionary algorithm with a novel deep ensemble reinforcement learning strategy for COVID-19 diagnosis
Appl. Soft Comput.
(2021) - et al.
Scalable deep learning-based recommendation systems
ICT Express
(2019) - et al.
A social recommender mechanism for e-commerce: Combining similarity, trust, and relationship
Decis. Support Syst.
(2013) - et al.
Deep sparse autoencoder prediction model based on adversarial learning for cross-domain recommendations
Knowl.-Based Syst.
(2021) - et al.
Collaborative filtering with a deep adversarial and attention network for cross-domain recommendation
Inf. Sci.
(2021)
An effective trust-based recommendation method using a novel graph clustering algorithm
Physica A
Data pre-processing evaluation for text mining: transaction/sequence model
Procedia Comput. Sci.
TCARS: Time-and community-aware recommendation system
Future Gener. Comp. Syst.
Characterizing context-aware recommender systems: A systematic literature review
Knowl.-Based Syst.
Collaborative filtering and deep learning based recommendation system for cold start items
Expert Syst. Appl.
Trust aware recommender system using swarm intelligence
Journal of computational science
Improving cluster analysis by co-initializations
Pattern Recogn. Lett.
A deep learning based trust-and tag-aware recommender system
Neurocomputing
Modeling trust-aware recommendations with temporal dynamics in social networks
IEEE Access
Enhanced arabic information retrieval: Light stemming and stop words
AutoTrustRec: Recommender system with social trust and deep learning using autoEncoder
Multimedia Tools Appl.
Latent Dirichlet allocation
J. Mach. Learn. Res.
Wide & deep learning for recommender systems
Cross-domain recommendations without overlapping data: Myth or reality?
Cross-domain recommender systems
On deep learning for trust-aware recommendations in social networks
IEEE Trans. Neural Networks Learn. Syst.
Adaptive subgradient methods for online learning and stochastic optimization
J. Machine Learn. Res.
Using gamification to tackle the cold-start problem in recommender systems
Beyond the stars: improving rating predictions using review text content
Cross-domain Recommendation with Bridge-Item Embeddings
ACM Trans. Knowledge Disc. Data (TKDD)
POI recommendation: Towards fused matrix factorization with geographical and temporal influences
DeepFM: a factorization-machine based neural network for CTR prediction
Neural factorization machines for sparse predictive analytics
A general cross-domain recommendation framework via Bayesian neural network
Neural collaborative filtering
Initializing matrix factorization methods on implicit feedback databases
J. Univers. Comput. Sci.
Collaborative metric learning
Conet: Collaborative cross networks for cross-domain recommendation
Cited by (4)
A Multi-User-Multi-Scenario-Multi-Mode aware network for personalized recommender systems
2024, Engineering Applications of Artificial IntelligenceAdditive autoencoder for dimension estimation
2023, NeurocomputingAsCDPR: a novel framework for ratings and personalized preference hotel recommendation using cross-domain and aspect-based features
2023, Data Technologies and Applications
Adeel Ahmed received his MPhil degree in computer science from Quaid-i-Azam University, Islamabad, Pakistan, in 2011. He is now PhD student at department of Computer Science, Quaid-i-Azam University, Islamabad, Pakistan. He was with the software industry for some years. His research interests include Machine learning, Recommendation systems, Social Network Analysis and Information Visualization.
Khalid Saleem received the MSc degree in computer science from Quaid-i-Azam University, Pakistan, in 1994 and the MPhil and PhD degrees in computer science from the University of Montpellier 2, France, in 2005 and 2008, respectively. He was with the software industry for some years. He currently works as an Assistant professor in Quaid-i-Azam University, Pakistan. He is a president of PAK-France Alumni Network. His research interests include Machine learning, Deep learning, Schema matching and integration, Data Analytics, Stegnography,Cryptography and Recommender Systems.
Osman Khalid has completed his PhD from North Dakota State University, USA and his masters from Center for Advanced Studies in Engineering. He is currently Assistant Professor in COMSATS University Islamabad, Abbottabad Campus, Pakistan. His research areas include: Wireless Networks, Network Routing Protocols, Internet of Things, and Fog Computing. His website is: http://osman.pakproject.com.
Jiechao Gao received the BS degree in Jilin University 2016, and the MS degree in Columbia University in the city of New York 2018. He is currently a Ph.D. student in the Department of Computer Science of University of Virginia. His research interests include distributed networks, cloud computing, IoT, machine learning algorithms and applications.
Umer Rashid received his BS(CS) degree in computer science from University of Lahore, Pakistan, in 2005. He received his MPhil and PhD degrees in computer science from Quaid-i-Azam University, Islamabad, Pakistan in 2008 and 2017, respectively. He has teaching and research experience in international organizations. Currently he is serving as an Assistant Professor in Quaid-i-Azam University, Islamabad, Pakistan. His research interests include user centered computing, multimedia information retrieval, and multimedia technology. His work is published in international journals and conferences.